Re: [RFC] mm,tracing: improve current situation

From: Vlastimil Babka
Date: Tue Apr 08 2014 - 11:44:18 EST


On 04/04/2014 01:24 AM, Sasha Levin wrote:
On 04/03/2014 05:44 PM, Davidlohr Bueso wrote:
Hi All,

During LSFMM Dave Jones discussed the current situation around
testing/trinity in the mm. One of the conclusions was that basically we
lack tools to gather the necessary information to make debugging a less
painful process, making it pretty much a black box for a lot of cases.

One of the suggested ways to do so was to improve our tracing. Currently
we have events for kmem, vmscan and oom (which really just traces the
tunable updates) -- In addition Dave Hansen also also been trying to add
tracing for TLB range flushing, hopefully that can make it in some time
soon. However, this lacks the more general data that governs all of the
core VM, such as vmas and of course the mm_struct.

To this end, I've started adding events to trace the vma lifecycle,
including: creating, removing, splitting, merging, copying and
adjusting. Currently it only prints out the start and end virtual
addresses, such as:

bash-3661 [000] .... 222.964847: split_vma: [8a8000-9a6000] => new: [9a6000-9b6000]

Now, on a more general scenario, I basically would like to know, 1) is
this actually useful... I'm hoping that, if in fact something like this
gets merged, it won't just sit there. 2) What other general data would
be useful for debugging purposes? I'm happy to collect feedback and send
out something we can all benefit from.

I think that adding more tracepoints might be more useful for debugging performance-related problems (e.g. compaction) that don't manifest as panic, and that VM_BUG_ON is more suited for this kind of debugging. But I might be wrong.

There's another thing we have to think about, which is the bottleneck of
getting that debug info out.

Turning on any sort of tracing/logging in mm/ would trigger huge amounts
of data flowing out. Any attempt to store that data anywhere would result
either in too much interference to the tests so that issues stop reproducing,
or way too much data to even be able to get through the guest <-> host pipe.

I was working on a similar idea, which is similar to what lockdep does now:
when you get a lockdep spew you see a nice output which also shows call
traces of relevant locks. What if, for example, we could make dump_page()
also dump the traces of where each of it's flags was set or cleared?

Hm doesn't the oops printing already print accumulated trace buffers? Wouldn't it be easier to post-process that instead of trying to do some smart "unwinding" during oops? Is it possible to enable tracing without actually consuming the data, just for this purpose?

Vlastimil

Thanks,
Sasha

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/